Mixed models in the validation of satellite derived data for hydrological studies

TEICH, I. ; GARCÍA, C.L. ; BUGAN, R.D.H. ; JOVANOVIC, N. ; BALZARINI, M.

Florencia

Conferencia; XXVIIth International Biometric Conference; 2014

International Biometric Society

A wealth of Earth Observation (EO) products obtained from remotely sensed image data covering large areas is now available to the scientific community. These products include water cycle-related variables such as evapotranspiration, soil moisture and vegetation indexes at high spatial and temporal resolutions, which can be used for the monitoring of natural ecosystems and crops as well as other operational and management actions that are crucial in the global change scenario. Linear Mixed Models (LMM) offer a flexible framework for studying spatially and temporally correlated data, such as the provided by satellite images, and consequently they can be used to validate EO products. In this work we adjusted several LMM to associate satellite derived evapotranspiration and vegetation indexes to variables obtained from distributed hydrological models (JAMS and J2000) implemented in the Sandspruit River Catchment (South Africa) and in the San Antonio River Catchment (Argentina). In the first case, data corresponds to 1660 time series, each belonging to a Hydrological Response Unit (HRU), spanning from 2009 to 2011. In the Argentinean catchment, 4271 HRUs were analyzed with data spanning from 1998 to 2012. Alternative LMM accounting for temporal and spatial autocorrelation were compared using the Akaike Information Criteria and classic Likelihood Ratio Tests. Temporal correlation was taken into account including HRUs as a random effect and specifying the correlation of errors with autorregressive structures. The spatial correlation between observation pairs was considered through the modeling of the residual covariance matrix (Exponential, Gaussian and Spherical) and including spatial data in the mean structure. Selected models were validated using data from the most recent year and relative errors were calculated. Accounting for temporal and spatial autocorrelation with LMM proved to be useful to obtain more reliable and accurate predictions of soil water content and evapotranspiration estimations. Appropriate usage, awareness of limitations and correct interpretation of remote sensing data can facilitate water management and planning operations.